Resnet only use contiguous in training mode. #12977
Conversation
Signed-off-by: jiqing-feng <[email protected]>
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
The failing test passes on But with this PR, that fails. Could you please check? |
Signed-off-by: jiqing-feng <[email protected]>
|
The behaviour change has little impact on some models' precision; just changing the tolerance from |
Signed-off-by: jiqing-feng <[email protected]>
|
But we need the tests to pass on our CI. |
|
I will check these failed tests. |
|
Hi @sayakpaul . The failed test seems network issue. Please check it. Thanks! |
|
Hi @sayakpaul . Please retrigger the CI. Thanks! |
|
It's so weird. The CI already passed in commit aa65c5c, but failed after update branch. @sayakpaul . Do you have any clue about it? |
Hi @sayakpaul . Would you please re-run the CI to see if the error disappears? Thanks! |
|
I see the lora tests also failed in #13003 |
| @require_hf_hub_version_greater("0.26.5") | ||
| @require_transformers_version_greater("4.47.1") | ||
| def test_save_load_dduf(self, atol=1e-4, rtol=1e-4): | ||
| def test_save_load_dduf(self, atol=1e-3, rtol=1e-3): |
There was a problem hiding this comment.
We shouldn't change it at the main test level. If a model test failing, we should override the method at the corresponding test class and relax tolerance.
3 participants
As the contiguous may have a negative impact on CPU performance, ResNet can only use it in training mode.
Fix #12975